Super-resolution of car plate images using generative adversarial networks

Abstract

Car plate recognition is used in traffic monitoring and control systems such as intelligent parking lot management, finding stolen vehicles, and automated highway toll. Car plate recognition consists of several stages of processing namely, car plate localization, extraction, and recognition which consists of Optical Character Recognition (OCR). However, in practice, Low-Resolution (LR) images or videos are widely used in surveillance systems. In low resolution surveillance systems, the car plate text is often illegible. Other than that, small car plate due to the distance and illumination cause the car plate recognition to fail as well. Super-Resolution (SR) techniques can be used to improve the car plate quality by processing a series of LR images into a single High-Resolution (HR) image. Today, the best upscaling algorithms cannot effectively reconstruct data that does not exist. Recovering the HR image from a single LR is still an ill-conditioned problem for SR. Previous methods always minimize the mean square loss in order to improve the peak signal to noise ratio(PSNR). However, minimizing the mean square loss leads to overly smoothed reconstructed image. In this project, Generative Adversarial Networks (GANs) based SR is proposed to reconstruct the LR images into HR images. Besides that, perceptual loss is proposed to solve the smoothing issue. The quality of the GAN based SR generated images will be compared to existing techniques such as bicubic, nearest and Super-Resolution Convolution Neural Network (SRCNN). The results show that the reconstructed images using GANs based SR achieve better results in term of perceptual quality compared to previous methods.